What is jitter? How packet loss and MOS affect voice quality
Jitter is the uneven arrival of voice packets on an internet call: instead of landing at a steady beat, some show up early, some late, and some too late to play. On a human call, the ear rarely notices the small gaps. On a call to an AI agent, those gaps become missing syllables in the transcript, and the model acts on what it thinks it heard.
Picture the claims line. A caller reads their policy number, but a jitter spike drops a fragment of the third digit. The AI agent confirms a number the caller never said, the caller repeats themselves, and the call escalates to a human. The operations lead checks the network dashboard the next morning: average latency is inside target, the carrier’s monthly service level agreement (SLA) report is green, and no ticket is open. Containment, however, is down month over month.
That mismatch, healthy averages alongside failing calls, is what per-call measurement exposes. The rest of this guide covers what jitter is, how it interacts with latency and packet loss, why AI agents need a stricter voice quality floor than human agents, and how to write that floor into the contracts that govern the trunks.
What is jitter in VoIP?
Jitter is the variation in the time between arriving voice packets on a Voice over Internet Protocol (opens in a new tab) (VoIP) call. Speech travels as a stream of small packets, and each one is meant to arrive at a steady interval, so playback sounds continuous.
On a real network, some packets arrive early and some late. The receiving device holds them in a jitter buffer and releases them at an even pace to keep playback smooth.
The buffer has limits. The standards body approved Recommendation P.940 (opens in a new tab) in March 2025, which states that receiving devices translate jitter into extra delay or packet loss. When a packet misses its slot, the device either holds it, which adds delay, or discards it entirely.
A call with fast average delivery can therefore sound broken when arrivals are irregular. That coupling between arrival variation and audible quality is why jitter cannot be judged in isolation from the two metrics network teams report alongside it: latency and packet loss.
Why missing packets compound across a call
Packet loss is the share of voice packets that never reach the receiving device in time to be played. A voice call, unlike a web page load, cannot pause and request a missing piece of data: the moment for playback passes, and the slice of speech is gone.
The FCC’s Measuring Broadband America program (opens in a new tab) reports that packet loss of a few tenths of a percent is common and is unlikely to affect the perceived quality of most Internet applications, but voice sits at the strict end of that scale because the 1% standard for packet loss is commonly accepted as the point at which highly interactive applications such as VoIP experience significant degradation in quality.
Loss can also arrive indirectly when jitter overruns the receiving buffer, and the device discards late packets. Judging its effect on a caller means combining it with delay and codec behavior into a single number.
What MOS measures
A mean opinion score (MOS) is a single number that expresses how a call sounded by compressing delay, loss, codec distortion, echo, and noise into one value. The label reflects its origin: listening subjects rate real calls on a five-point scale running from excellent through good, fair, and poor to bad, and the score is the mean of those ratings.
Panels do not scale to production, so most monitoring tools estimate MOS with the E-model in Recommendation G.107 (opens in a new tab), which estimates transmission quality for voice calls as a function of roughly two dozen parameters, including network-layer performance measures, background noise at the sender and receiver, electric circuit noise, and parameters specific to the voice codec; tools that also capture the audio use Recommendation P.863 (opens in a new tab).
Every version of the score describes what a human listener would report, which is why the three network inputs feeding it deserve inspection on their own terms.
How jitter differs from latency and packet loss
Network reports carry three separate voice metrics, and each one describes a different failure mode. Reading them together prevents the situation where average one-way latency sits well inside target while a single trunk repeatedly overflows its jitter buffer during a busy hour.
The table below summarizes how each metric behaves, and the working ceilings network engineering guidance commonly cites for voice traffic.
Metric | What it measures | Effect on the caller | Commonly cited ceiling |
|---|---|---|---|
Latency | The one-way time a single packet takes from speaker to listener | Stretches the pause (opens in a new tab) between one party finishing and the other hearing it | 150 ms or less |
Jitter | The variation in arrival time between consecutive packets | Forces the buffer either to grow, adding delay, or to drop late packets | 30 ms or less |
Packet loss | The share of packets that never reach the receiving device in time to be played | Removes a slice of speech that cannot be recovered before playback | 1% or less |
Cisco Meraki’s guidance for wireless voice recommends a target of 10 ms of jitter and no more than 50 ms, reinforcing that the working ceiling for AI-handled calls should sit well below the general VoIP figure. An averaged jitter number across a month cannot show whether any single call crossed those lines.
Latency alone costs the conversation time; jitter and loss cost it content. Missing content, however, does not stay contained to the moment it happened, which is why network teams also have to watch how packet loss accumulates across a full call.
Why AI agents need a higher voice quality floor than human agents
Human-to-human calls can remain usable at a lower MOS because people repair conversations as they go. A human agent who half-hears a policy number says, “Sorry, could you repeat that?” and the caller does. An AI agent acts on the transcript it received. The contrast between conversational repair and transcript-based action creates a need for a stricter, consistently maintained voice-quality target for AI calls.
The margin is wider because automatic speech recognition (opens in a new tab) (ASR) accuracy can decline as audio degrades. Human listeners can use context and ask for repetition. A listener fills in a clipped syllable from context; a recognition model scores the acoustic evidence it has and commits. A recognition model alone has no conversational repair step, so each missing audio fragment can become a wrong word in the transcript instead of prompting a clarifying question.
Word error rate (opens in a new tab) (WER), the share of words transcribed incorrectly, can climb even when loss counts still sit comfortably inside the human floor. A degraded signal reaches the model through several distinct failures, and each one is invisible on the network dashboard:
Garbled transcription: a lost or late packet removes part of a word, and the ASR emits its best guess instead of a blank. The large language model (LLM) downstream never sees the uncertainty.
False endpointing: a jitter spike creates a silence in the audio stream that voice activity detection reads as the caller finishing. The AI agent answers a half-sentence, and the caller starts over.
Missed entities: a policy number, date of birth, or postcode transcribed with one wrong digit fails authentication or opens the wrong record, and the caller has no way to know which digit went missing.
Wrong intent: a misheard verb changes the request, so a call about canceling a payment is routed as a call about making one.
A degraded trunk raises escalations and lowers containment even when network ticket volume stays unchanged.
How to assign ownership for voice quality
A voice call to an AI agent crosses multiple vendors’ equipment: the carrier that delivers it, the contact center as a service (CCaaS) platform that queues it, the speech-to-text (STT) engine that transcribes it, the LLM that decides what to do, and the text-to-speech (TTS) engine that answers. When quality degrades, each vendor can truthfully report that its component performed to specification, so the organization must name the owner of the whole conversation in writing.
A VoIP SLA can follow this pattern when the carrier commits to jitter as a monthly average calculated from data it collects periodically across a network. A monthly average absorbs a busy-hour spike that ruined an hour of calls, so the carrier stays within SLA on the same afternoon the AI agent’s containment falls. Governance replaces that arithmetic with decisions your team can write into a contract or a monitoring rule:
Measurement granularity: specify jitter and loss per call, sampled continuously, with the raw distribution available to the buyer, and exclude monthly means as the compliance measure.
AI-specific MOS floor: write the AI vendor’s production floor into the carrier and CCaaS agreements.
Escalation trigger: define the alert that fires when per-call MOS degradation and a containment drop coincide, and name the team that owns the joint investigation.
Simulation under degraded conditions: before go-live, run the AI agent against audio with injected jitter, loss, and background noise and record where authentication and intent accuracy start to fall.
Per-call monitoring has to work at production volume. HSE runs 600 simultaneous calls (opens in a new tab) and 3 million automated calls annually, and at that scale, per-call MOS is the only view where a spike on one trunk shows up before the containment number does. That visibility gives the named owner time to investigate the failing route before the monthly carrier report hides it.
Set the voice quality floor before your AI agent takes calls
Voice is the channel customers reach for when the stakes are highest: an urgent claim, a payment dispute, an account they cannot access. An AI agent answering those calls has to act on the audio it receives, without the conversational repair a human colleague would use to recover a missing digit or a clipped word. That makes the voice quality floor a sales and support decision as much as a network one: containment, authentication, and first-contact resolution all rest on the transcript the model actually sees, and the contract governing the trunks is where that floor becomes enforceable.
Parloa (opens in a new tab) is an AI Agent Management Platform that supports lifecycle management across Build, Optimize, and Observe, with conversation simulation under degraded audio before launch and continuous monitoring across 140+ languages. Its certifications include ISO 27001:2022, ISO 17442:2020, ISO/IEC 42001:2023, SOC 2 Type 1 & 2, PCI DSS, HIPAA, GDPR, DORA, and the EU AI Act.
Book a demo (opens in a new tab) to see how voice quality monitoring protects AI agent containment in production.
FAQs about jitter and voice quality
What is an acceptable jitter level for VoIP?
Network engineering guidance sets a firm ceiling on jitter for voice traffic. For calls that an AI agent answers, set the target below the general VoIP figure, because a speech recognition engine alone does not ask the caller to repeat a word the buffer dropped. Treat the published ceiling as an upper bound and negotiate the AI target separately.
How do you reduce jitter on contact center voice traffic?
Mark voice packets with quality of service (QoS) tags so routers and switches prioritize them over bulk data. Tune the jitter buffer on the receiving side so it absorbs normal variation without adding delay that the conversation can feel. Where traffic crosses the public internet, software-defined WAN (SD-WAN) path selection routes voice over the link showing the least variation.
What causes network jitter?
Congestion is one cause: packets queue behind other traffic at a busy router and leave at uneven intervals. Route changes send consecutive packets down paths of different lengths, wireless interference forces retransmissions on Wi-Fi and cellular hops, and misconfigured buffers add their own variation. Monitoring each trunk separately helps the network team isolate the cause before it affects more calls.
What is a good MOS score for a contact center?
A good score depends on who is listening. Human-to-human calls tolerate a fair-to-good score because both parties can repair misunderstandings as they go. Calls handled by an AI agent need a good-to-excellent score, because the transcript the model receives cannot be repaired after the fact.
Can a carrier be within its SLA while individual calls sound bad?
Yes. A carrier commitment based on a monthly average can absorb a severe spike lasting an hour, so the spike disappears into the mean. Only per-call or per-trunk measurement, written into the contract, shows the spike the caller experienced.
:format(webp))
:format(webp))